Abstract
Determining RNA secondary structures is a fundamental challenge in computational biology and molecular sensing. Experimental techniques such as X-ray crystallography, nuclear magnetic resonance, and cryo-electron microscopy can reveal RNA structures with atomic precision, but their high cost and time consuming nature limit large-scale applications. To address this issue, we introduce the Structure-Sensing Nucleotide Attention Learning framework (NTFold), a virtual sensing framework based on deep learning for accurate RNA secondary structure prediction. NTFold integrates a Nucleotide Attention Module (NAM) to explicitly model dependencies among nucleotides, thereby capturing fine-grained sequence correlations. The resulting correlation map is subsequently refined by a Structural Refinement Module (SRM), which preserves hierarchical spatial information and enforces structural consistency. Through this two stage learning paradigm, NTFold produces high-precision contact maps that enable reliable RNA secondary structure reconstruction. Extensive experiments demonstrate that NTFold outperforms existing deep learning-based predictors, highlighting its capability to learn both local and global nucleotide interactions in an sensor inspired manner. This study provides a new direction for integrating attention driven correlation modeling with structure-sensing refinement toward efficient and scalable RNA structural sensing.